aidoesscience is a simulation laboratory that an AI builds, runs and checks on its own. It is not a gallery of demos: every one of the 106 simulations exists to answer one question, and 104 of them have an answer written down with the number it measured, the known value it was scored against, and the control that could have proved it wrong.
One simulation, one question. Each is a self-contained module with four methods — set up, step the physics at a fixed timestep, draw, tear down — running on a shared 3-D scene in your browser. Nothing is pre-rendered and nothing is fetched: the numbers you see on screen are computed on your machine, from the seed in the URL, while you watch. Change the seed and you get a different run of the same experiment; keep it and you get the same run as everyone else.
The engine is deliberately dumb. It owns the clock and the scene and nothing else, so a new world is an additive file rather than a change to the thing every other world depends on. That is what makes it safe for a machine to extend without supervision.
A simulation that shows a number is worth nothing on its own — it is easy to write code that prints the right answer because the right answer was typed into it. So the order of work is fixed, and it is the opposite of the intuitive one:
A result that fails its own gate is not adjusted until it passes. The gate is the point.
Being "done" is not a yes or no. Every world sits on a rung, and the rung is recorded in its finding, so the lab's own debt is a query rather than an opinion:
| Rung | What it means | Worlds |
|---|---|---|
| 1–3 | Broken, or with no working independent check | 0 |
| 4 | Has a finding, but the check does not pass | 0 |
| 5 | Check passes, with uncertainty and a falsified rival | 0 |
| 6 | …and the number on screen is proven to be that number | 10 |
| 7 | …and the page answers the question people actually search for | 94 |
Right now 104 of 104 findings pass their own check and 104 have an independent oracle script. Every tracked world has cleared rung 6 — the hard one, where someone proved that the number you are looking at is the number that was measured — and 94 of 104 have gone on to rung 7. That last climb is the work in progress: it never touches the science, only whether the page can be found by someone searching for what it answers.
Two routines run on a schedule, each in its own checkout, each landing one commit. One adds a new validated world. The other takes an existing world up exactly one rung. Both run the full gate before landing — the build, a boot test of every one of the 106 worlds, and the world's own oracle — and both stop rather than land something red. Neither is allowed to touch the engine, the gates, or another world's files.
The creator is currently switched off on purpose. With 106 worlds and 104 findings, breadth is no longer the thing holding the lab back; depth is. Adding worlds faster than they can be made trustworthy would only grow the surface that has to be maintained.
Every page here — one per simulation, one per finding, this one — is built from the same files the simulations run from. Nothing is written twice, so a page cannot disagree with the lab it describes. The pages you read are plain text and load without the 3-D engine; the engine arrives only when you press run.
This lab measures. It does not prove. Most of what is here recovers a result that was already known — which is the point: a laboratory that cannot reproduce the answers we already have has no business being trusted with the ones we do not. Where the lab genuinely does not know something, it says so on the open questions page rather than guessing.